When Intel introduced its Xeon 6 platform earlier this year, the company said that the new CPUs will support multiplexed rank dual inline memory modules (MRDIMMs) with the promise of bringing together high data transfer rates, low latency, high capacity, and predictable power consumption. Late last week the company decided to remind the industry about its MRDIMM support probably because server makers are beginning to roll out systems with such memory modules. For example, NEC recently announced its soon-to-be-built supercomputer that will use MRDIMMs.
As server CPUs ramped up core counts in recent years, memory bandwidth available per core has decreased - despite the fact that DDR5 has greatly improved data transfer rates compared to DDR4, which created performance bottlenecks for compute-heavy workloads. MRDIMMs are designed to improve memory performance by effectively doubling available bandwidth.
MRDIMMs operate two DDR memory ranks in a multiplexed mode. These modules include not only additional memory devices but also an MRCD chip, which enables simultaneous access to both memory ranks, and MDB chips that handle multiplexing and demultiplexing. This configuration allows the CPU to communicate with MRDIMM modules at an impressive 8,800 MT/s transfer rate with Intel's Xeon 6 CPU. However, actual memory device components run at half this speed, which helps reduce latencies and power consumption. Intel's Xeon 6 processors support up to 3TB of MRDIMM memory at DDR5-8800 speed.
According to Intel and Micron, a 128GB DDR5-8800 MRDIMM provides up to 40% lower latency under load compared to a 128GB DDR5-6400 RDIMM, which greatly enhances real-world memory performance. Additionally, lowering the power consumption of memory using MRDIMMs is crucial as the energy requirements of memory modules in server systems can be on par with or even exceed the power consumption of certain server CPUs.
One of MRDIMM's main advantages is its seamless integration. It maintains the same connector and physical layout as RDIMMs, requiring no changes to the motherboard or physical configuration in servers. This compatibility means that server makers can easily integrate MRDIMMs without making any adjustments to their designs. In fact, even operators of datacenters can upgrade their Xeon 6 servers with MRDIMMs without making any adjustments to their infrastructure or software. Also, MRDIMMs retain all the reliability, availability, and serviceability features of DDR5 RDIMMs, including error correction.
Intel says it had collaborated extensively with memory vendors and industry partners to make MRDIMM an open standard. They contributed MRDIMM's specifications to the JEDEC standards organization in late 2022, and major memory vendors have already begun introducing MRDIMMs to the market, with more expected to follow..
Intel's senior engineer, George Vergis, who led the development of MRDIMMs, has also contributed to previous standards like DDR5. Vergis and his team started working on MRDIMM in 2018 and developed prototypes by 2021, achieving proof of concept before collaborating with the industry to finalize MRDIMM as an open standard.
